Short answer: to track brand mentions in ChatGPT, write a fixed set of buyer questions that contain no brand names, run them in ChatGPT with the same settings on a regular cadence, detect every brand in each answer with exactly the same rules, and record mentions, recommendations and citations as separate columns, keeping the answer text as evidence. Setting this up takes an afternoon. Keeping it honest is the part that takes discipline, because every shortcut makes the number look better than it is.
The steps below are the rules we apply to our own measurement. Several exist because we broke them first, and the incidents, with their dates and the numbers they produced, are described where they apply. None of this requires special software: a spreadsheet and a calendar reminder will do for a first quarter. What it requires is that the rules do not bend when the number is disappointing.
Before you track brand mentions in ChatGPT, check it can read you
Tracking mentions of a brand that ChatGPT cannot fetch will mostly tell you about other people's pages. OpenAI documents three agents with different jobs: OAI-SearchBot builds the index behind ChatGPT's search features, ChatGPT-User fetches pages a user asks about, and GPTBot collects content that may be used for training. Blocking all three with one rule is the most common way sites remove themselves from answers, and the agent most often caught by accident is OAI-SearchBot, blocked by sites that only meant to decline training.
Check the first two against OpenAI's bot documentation before you read anything into a low mention rate. Then check that your server actually answers them: a permissive robots.txt in front of a CDN rule that returns 403 to those agents is a block in practice, and the file will keep saying otherwise. A low number on a site that ChatGPT cannot read is a crawling problem, not a content problem, and no amount of tracking will fix it.
Step 1: build a question set without brand names
The questions should be the decisions buyers actually ask an assistant to help with, written as full sentences with a situation in them: a budget, a constraint, a use case. Assistants answer situations, and a question with a constraint produces a different set of named brands from the bare category term.
| Intent | Example question (illustrative) | What being absent here usually means |
|---|---|---|
| Research | What should a small clinic look for in patient scheduling software? | The category is being explained without you; your explanatory content is missing or not quotable |
| Comparison | Which scheduling tools work for a clinic with three locations and shared staff? | You are not in the shortlist the assistant builds; check the round-ups and review sites it cites |
| Purchase | What is the best scheduling software for a dental practice that needs SMS reminders? | You lose at the moment of decision; the facts that answer the constraint are not on your pages |
Record the intent on every question. Being named in research questions and absent from purchase questions is a different problem, with a different fix, from the reverse, and one overall percentage cannot tell them apart.
The trap is the self-referential question. If a question names your brand, the answer will too, and that is an echo rather than visibility. A model asked to generate a question set drifts towards these on its own: one of ours once produced fifteen questions of which thirteen contained the brand name, and the resulting score read 13% across five engines when the honest value was zero. Competitor names are excluded for the same reason. Filter them when the questions are generated, not only when they are scored, or you spend quota on questions you will throw away. Building a prompt set covers the whole process.
Step 2: fix the settings and the cadence
Everything that changes the measurement has to be recorded with each run: which engine and model answered, whether web search was on, the question set version and the date. Enabling web search once moved a site from 3% to 13% while the site did nothing. Stamp every run with a methodology version, and treat runs on different versions as separate series rather than as a trend.
Keep the question set fixed between runs unless you change it deliberately. Every percentage has the question set as its denominator; add easier questions and the average rises while nothing improves. When you do change it, restate the history instead of joining the old line to the new one.
Choose a cadence you can afford to keep. Our own monitoring defaults to monthly, because answer runs come from a shared account-level quota and a weekly default would quietly spend a team's whole month on monitoring. Whatever you choose, check that every completed run actually reaches the series: we once found that runs taking longer than 10 minutes were never snapshotted, so the trend silently left them out. A trend that skips the slow runs is biased towards whatever makes runs fast, which is rarely what you meant to measure.
Step 3: detect every brand the same way
Count every brand an answer names, not only the ones on your list. Buyers see whatever the assistant names, including brands you have never heard of. Then apply the same rules to everyone, yourself included:
- One code path for all brands. Matching your own name loosely and competitors only against a fixed list inflates your share by construction. Asymmetric detection once showed a brand at 31% share of voice and rank two; measured symmetrically it was 0% and last.
- Resolve spelling variants before counting. Casing, punctuation, legal suffixes and a product line named instead of the company each split one brand into several rows. If "Acme" and "Acme Systems" fold into one entity for you, they fold for the competitor too. Brand normalisation explains the mechanics.
- Show the number of brands next to your share. Being 1% of 117 brands is a different situation from 1% of five, and the percentage alone cannot tell them apart.
- Withhold share when the sample is thin. A percentage from a handful of answers is noise wearing the costume of a signal, and it will swing on the next run.
Negative mentions belong in the count of brands named, because a buyer reads them, but they do not belong in your wins. An answer that names you in order to steer buyers elsewhere has shown them your name and told them not to choose you; scoring it as a point in your favour reverses its meaning. How to choose and maintain the competitor list itself is covered in competitor set design.
Step 4: record each ChatGPT brand mention in separate columns
"Acme is often criticised for slow support, so most buyers choose another tool" contains your brand. Whether that counts depends on which question the number is answering. Keep these as separate columns for every answer:
| Column | What it records | The question it answers |
|---|---|---|
| Mentioned | The brand is named anywhere in the answer | Do buyers see our name at all? |
| Recommended | The answer actively suggests the brand for the question asked | Are we the answer, or a footnote? |
| Cited | Your domain is among the sources marked in the answer, resolving to a real URL | Is our content trusted as evidence? |
| Sentiment | Whether the mention is positive, neutral or negative | What do buyers learn about us? |
Your page can be cited as a source while the answer recommends a competitor. Merging the columns flatters the report and hides that gap. For citations, count only the sources marked inside the answer text: an assistant typically retrieves a dozen or more pages and cites a few, and a marker that does not resolve to a real URL should be discarded rather than counted. Brand share in AI answers goes further into how share should be counted.
Keep the evidence, and alert only on changes that repeat
Store the answer, not just the score
Scores summarise; evidence explains what to do. Store the question, the full answer, the engine, the settings, the date, the cited URLs and the other brands named, for every result.
Without the answer text you cannot tell a detection bug from a real change, and you will see both. Store failed runs as failed, not as zero: a model that timed out produced no evidence either way, and rendering it as 0% invents a decline that never happened.
Alert on transitions that hold for two runs
ChatGPT's answers vary between runs even when nothing has changed. Two rules keep alerts useful. Fire only when the same change appears in two consecutive runs, and fire on the transition rather than on the state, so a condition that persists does not repeat the same alert every run and bury the new ones. If the previous run used a different methodology version, treat it as having no previous run at all. Alerting and thresholds sets out sensible defaults.
The whole method on one page
| Stage | The shortcut | What it does to the number | The defensible version |
|---|---|---|---|
| Questions | Let a model write them, unchecked | Brand-named questions echo you back; 13% instead of zero | Brand-free, full sentences, intent recorded, filtered at generation |
| Settings | Change web search or the model mid-series | A method change looks like progress; 3% to 13% with no work done | Methodology version on every run; no comparison across versions |
| Detection | Loose match for you, fixed list for rivals | Share inflated by construction; 31% instead of 0% | One rule for every brand, variants resolved first |
| Columns | One "visible" flag | Criticism and citations counted as wins | Mention, recommendation, citation and sentiment kept apart |
| Failures | Timed-out runs stored as 0% | An invented decline and a false alert | Failed runs marked as no data |
If you would rather not run this by hand, InsightWonder's prompt monitoring applies these rules to scheduled ChatGPT runs and keeps the answer text beside every number.
Sources
- OpenAI: overview of OpenAI crawlers (GPTBot, OAI-SearchBot, ChatGPT-User)
- The incidents cited in this article (the brand-named question set, asymmetric share of voice, the method change and the missing snapshots) are from InsightWonder's own measurement records, dated 2026-07-20 to 2026-09-07.